IMPORT LIBRAIRIES

library(phyloseq) # for phyloseq object
library(ggplot2)
library(cowplot)
library(plyr)
library(dplyr)
library("plotly") # plot 3D
library("microbiome") # for centered log-ratio
library("coda") # Aitchison distance
library("coda.base") # Aitchison distance
library("vegan") # NMDS
library(pheatmap) # for heatmap
library(dendextend) # for dendrogram

IMPORT DATA

# Set path
path <- "~/Projects/IBS_Meta-analysis_16S"

# Import phyloseq object
physeq.fukui <- readRDS(file.path(path, "phyloseq-objects/physeq_fukui.rds"))

# Sanity check
physeq.fukui
## phyloseq-class experiment-level object
## otu_table()   OTU Table:         [ 6863 taxa and 110 samples ]
## sample_data() Sample Data:       [ 110 samples by 9 sample variables ]
## tax_table()   Taxonomy Table:    [ 6863 taxa by 7 taxonomic ranks ]
## phy_tree()    Phylogenetic Tree: [ 6863 tips and 6861 internal nodes ]
## refseq()      DNAStringSet:      [ 6863 reference sequences ]

Phylogenetic tree was computed with the package phangorn, and the script was run on a cluster. Let’s check we have correctly generated a phylogenetic tree.

# Look at the tree
plot_tree(physeq.fukui, color = "host_disease", ladderize="left")

ABUNDANCES

1. Absolute abundances

# Plot Phylum
plot_bar(physeq.fukui, fill = "Phylum") + facet_wrap("host_disease", scales="free") +
  theme(axis.text.x = element_text(size = 8))+
  labs(x = "Samples", y = "Absolute abundance", title = "fukui dataset (2020)")

# Plot Class
plot_bar(physeq.fukui, fill = "Class")+ facet_wrap("host_disease", scales="free") +
  theme(axis.text.x = element_text(size = 8))+
  labs(x = "Samples", y = "Absolute abundance", title = "fukui dataset (2020)")

Sequencing depth characteristics of the fukui dataset:
- minimum of 3029 total count per sample
- median: 2.7533510^{4} total count per sample
- maximum of 4.940610^{4} total count per sample

2. Relative abundances

# Agglomerate to phylum & class levels
phylum.table <- physeq.fukui %>%
  tax_glom(taxrank = "Phylum") %>%                     # agglomerate at phylum level
  transform_sample_counts(function(x) {x/sum(x)} ) %>% # Transform to rel. abundance
  psmelt()                                             # Melt to long format

class.table <- physeq.fukui %>%
  tax_glom(taxrank = "Class") %>%
  transform_sample_counts(function(x) {x/sum(x)} ) %>%
  psmelt()


# Plot relative abundances
ggplot(phylum.table, aes(x = reorder(Sample, Sample, function(x) mean(phylum.table[Sample == x & Phylum == 'Firmicutes', 'Abundance'])),
                         y = Abundance, fill = Phylum))+
  facet_wrap(~ host_disease, scales = "free") + # scales = "free" removes empty lines
  geom_bar(stat = "identity") +
  theme(axis.text.x = element_text(size = 8, angle = -90))+
  labs(x = "Samples", y = "Relative abundance", title = "Fukui dataset (2020)")

ggplot(class.table, aes(x = Sample, y = Abundance, fill = Class))+
  facet_wrap(~ host_disease, scales = "free") + # scales = "free" removes empty lines
  geom_bar(stat = "identity") +
  theme(axis.text.x = element_text(size = 8, angle = -90))+
  labs(x = "Samples", y = "Relative abundance", title = "Fukui dataset (2020)")

3. Firmicutes/Bacteroidota ratio

# Extract abundance of only Bacteroidota and Firmicutes
bacter <- phylum.table %>%
  filter(Phylum == "Bacteroidota") %>%
  select(c('Sample', 'Abundance', 'host_disease', 'Phylum')) %>%
  arrange(Sample)

firmi <- phylum.table %>%
  filter(Phylum == "Firmicutes") %>%
  select(c('Sample', 'Abundance', 'host_disease', 'Phylum')) %>%
  arrange(Sample)

# Calculate log2 ratio Firmicutes/Bacteroidota
ratio.FB <- data.frame('Sample' = bacter$Sample,
                       'host_disease' = bacter$host_disease,
                       'Bacteroidota' = bacter$Abundance,
                       'Firmicutes' = firmi$Abundance)
ratio.FB$logRatioFB <- log2(ratio.FB$Firmicutes / ratio.FB$Bacteroidota)

# Plot log2 ratio Firmicutes/Bacteroidota
ggplot(ratio.FB, aes(x = host_disease, y = logRatioFB))+
  geom_boxplot(outlier.shape = NA)+
  geom_jitter(width=0.1)+
  labs(x = "",  y = 'Log2(Firmicutes/Bacteroidota)', title = "Firmicutes:Bacteroidota ratio")+
  theme_cowplot()

# Statistical test
wilcox.test(ratio.FB[ratio.FB$host_disease == "IBS","logRatioFB"],
            ratio.FB[ratio.FB$host_disease == "Healthy","logRatioFB"]) # p = 0.24
## 
##  Wilcoxon rank sum test with continuity correction
## 
## data:  ratio.FB[ratio.FB$host_disease == "IBS", "logRatioFB"] and ratio.FB[ratio.FB$host_disease == "Healthy", "logRatioFB"]
## W = 925, p-value = 0.2414
## alternative hypothesis: true location shift is not equal to 0

NORMALIZE DATA

# Sanity check no sample with less than 500 total count
table(sample_sums(physeq.fukui)<500) # all FALSE

#____________________________________________________________________
# PHYLOSEQ OBJECT WITH NON-ZERO COMPOSITIONS
physeq.NZcomp <- physeq.fukui
otu_table(physeq.NZcomp)[otu_table(physeq.NZcomp) == 0] <- 0.5 # pseudocounts

# Sanity check that 0 values have been replaced
# otu_table(physeq.fukui)[1:5,1:5]
# otu_table(physeq.NZcomp)[1:5,1:5]

# transform into compositions
physeq.NZcomp <- transform_sample_counts(physeq.NZcomp, function(x) x / sum(x) )
table(rowSums(otu_table(physeq.NZcomp))) # check if there is any row not summing to 1

# Save object
saveRDS(physeq.NZcomp, file.path(path, "data/analysis-individual/Fukui-2020/02_EDA-Fukui/physeq_NZcomp.rds"))

#____________________________________________________________________
# PHYLOSEQ OBJECT WITH RELATIVE COUNT (BETWEEN 0 AND 1)
physeq.rel <- physeq.fukui
physeq.rel <- transform_sample_counts(physeq.rel, function(x) x / sum(x) ) # divide each count by the total number of counts (per sample)

# check the counts are all relative
# otu_table(physeq.fukui)[1:5, 1:5]
# otu_table(physeq.rel)[1:5, 1:5]

# sanity check
table(rowSums(otu_table(physeq.rel))) # check if there is any row not summing to 1

# save the physeq.rel object
saveRDS(physeq.rel, file.path(path, "data/analysis-individual/Fukui-2020/02_EDA-Fukui/physeq_relative.rds"))

#____________________________________________________________________
# PHYLOSEQ OBJECT WITH COMMON-SCALE NORMALIZATION
physeq.CSN <- physeq.fukui
physeq.CSN <- transform_sample_counts(physeq.CSN, function(x) (x*min(sample_sums(physeq.CSN))) / sum(x) )

# sanity check
table(rowSums(otu_table(physeq.CSN))) # check that all rows are summing to the same total

# save the physeq.CSN object
saveRDS(physeq.CSN, file.path(path, "data/analysis-individual/Fukui-2020/02_EDA-Fukui/physeq_CSN.rds"))


#____________________________________________________________________
# PHYLOSEQ OBJECT WITH CENTERED LOG RATIO COUNT
physeq.clr <- physeq.fukui
physeq.clr <- microbiome::transform(physeq.fukui, "clr") # the function adds pseudocounts itself

# Compare the otu tables in the original phyloseq object and the new one after CLR transformation
otu_table(physeq.fukui)[1:5, 1:5] # should contain absolute counts
otu_table(physeq.clr)[1:5, 1:5] # should all be relative

# save the physeq.rel object
saveRDS(physeq.clr, file.path(path, "data/analysis-individual/Fukui-2020/02_EDA-Fukui/physeq_clr.rds"))

COMPUTE DISTANCES

1. UniFrac, Aitchison, Bray-Curtis and Canberra

First, let’s look at these four distances of interest.

#____________________________________________________________________________________
# Measure distances
getDistances <- function(){
  set.seed(123) # for unifrac, need to set a seed
  glom.UniF <- UniFrac(physeq.rel, weighted=TRUE, normalized=TRUE) # weighted unifrac
  glom.ait <- phyloseq::distance(physeq.clr, method = 'euclidean') # aitchison
  glom.bray <- phyloseq::distance(physeq.CSN, method = "bray") # bray-curtis
  glom.can <- phyloseq::distance(physeq.NZcomp, method = "canberra") # canberra
  dist.list <- list("UniF" = glom.UniF, "Ait" = glom.ait, "Canb" = glom.can, "Bray" = glom.bray)
  
  return(dist.list)
}


#____________________________________________________________________________________
# Plot in 2D the distances
plotDistances2D <- function(dlist, ordination="MDS"){
  plist <- NULL
  plist <- vector("list", 4)
  names(plist) <- c("Weighted Unifrac", "Aitchison", "Bray-Curtis", "Canberra")
  
  print("Unifrac")
  # Weighted UniFrac
  set.seed(123)
  iMDS.UniF <- ordinate(physeq.rel, ordination, distance=dlist$UniF)
  plist[[1]] <- plot_ordination(physeq.rel, iMDS.UniF, color="host_disease")
  
  print("Aitchison")
  # Aitchison
  set.seed(123)
  iMDS.Ait <- ordinate(physeq.clr, ordination, distance=dlist$Ait)
  plist[[2]] <- plot_ordination(physeq.clr, iMDS.Ait, color="host_disease")
  
  print("Bray")
  # Bray-Curtis
  set.seed(123)
  iMDS.Bray <- ordinate(physeq.CSN, ordination, distance=dlist$Bray)
  plist[[3]] <- plot_ordination(physeq.CSN, iMDS.Bray, color="host_disease")
  
  print("Canberra")
  # Canberra
  set.seed(123)
  iMDS.Can <- ordinate(physeq.NZcomp, ordination, distance=dlist$Can)
  plist[[4]] <- plot_ordination(physeq.NZcomp, iMDS.Can, color="host_disease")
  
  # Creating a dataframe to plot everything
  plot.df = plyr::ldply(plist, function(x) x$data)
  names(plot.df)[1] <- "distance"
  
  return(plot.df)
}

Now let’s plot!

# Get the distances & the plot data
dist.fukui <- getDistances()
plot.df <- plotDistances2D(dist.fukui)
## [1] "Unifrac"
## [1] "Aitchison"
## [1] "Bray"
## [1] "Canberra"
# Plot
ggplot(plot.df, aes(Axis.1, Axis.2, color=host_disease))+
  geom_point(size=6, alpha=0.5)  + scale_color_manual(values = c('blue', 'red'))+
  facet_wrap(distance~., scales='free', nrow=1)+
  theme_bw()+
  theme(strip.text.x = element_text(size=20))+
  labs(color="Disease")

# ggsave(file.path(path, "data/analysis-individual/PLOTS/plots-Fukui-EDA/distances4_MDS.jpg"), height = 4, width = 15)

2. Plot in 3D

For better visualization, we will also take a glance at reduction to 3D.

#____________________________________________________________________________________
# Plot 3D ordination
plotDistances3D <- function(d, name_dist){
  
  # Reset parameters
  mds.3D <- NULL
  xyz <- NULL
  fig.3D <- NULL
  
  # Reduce distance matrix to 3 dimensions
  set.seed(123)
  mds.3D <- metaMDS(d, method="MDS", k=3, trace = 0)
  xyz <- scores(mds.3D, display="sites") # pull out the (x,y,z) coordinates
  
  # Plot
  fig.3D <- plot_ly(x=xyz[,1], y=xyz[,2], z=xyz[,3], type="scatter3d", mode="markers",
                    color=sample_data(physeq.fukui)$host_disease, colors = c("blue", "red"))%>%
    layout(title = paste('MDS in 3D with', name_dist, 'distance', sep = ' '))
  
  return(fig.3D)
}

Now let’s plot!

plotDistances3D(dist.fukui$UniF, "UniFrac")
plotDistances3D(dist.fukui$Ait, "Aitchison")
plotDistances3D(dist.fukui$Canb, "Canberra")
plotDistances3D(dist.fukui$Bray, "Bray-Curtis")

HIERARCHICAL CLUSTERING

# For heatmaps: have group color
matcol <- data.frame(group = sample_data(physeq.fukui)[,"host_disease"])


# Function to get heatmap from the distances computed
plotHeatmaps <- function(dlist, fontsize){
  
  # Initialize variables
  i=1
  plist <- vector("list", 4)
  names(plist) <- names(dlist)
  
  # Loop through distances
  for(d in dlist){
    plist[[i]] <- pheatmap(as.matrix(d), 
                          clustering_distance_rows = d,
                          clustering_distance_cols = d,
                          fontsize = fontsize,
                          fontsize_col = fontsize-5,
                          fontsize_row = fontsize-5,
                          annotation_col = matcol,
                          annotation_row = matcol,
                          annotation_colors = list(host_disease = c('Healthy' = 'blue', 'IBS' = 'red')),
                          cluster_rows = T,
                          cluster_cols = T,
                          clustering_method = 'complete', # hc method
                          main = names(dlist)[i]) # have name of distance as title
    i <- i+1
  }
  
  return(plist)
}


# Get the heatmaps
heatmp.fukui <- plotHeatmaps(dlist = dist.fukui, fontsize = 8)

REPRODUCE PLOTS FROM PAPER

First, let’s try to reproduce figure 2A. It’s a PCoA of microbiome composition based on unweighted UniFrac.

# ___________________________________________________________________________________________________________
# FIGURE 1B
set.seed(123)
fig.2A <- ordinate(physeq.rel, "PCoA", "unifrac", weighted=FALSE)
plot_ordination(physeq.rel, fig.2A, color="host_disease", shape = "host_disease")+
  geom_point(size=5, alpha=0.5) +
  scale_color_manual(values = c('blue', 'red'))+ scale_shape_manual(values = c(17,16))+
  theme_classic()+
  ggtitle("PCoA on unweighted-UniFrac distance")

Now, let’s reproduce figure 2B. They performed a hierarchical clustering with Ward linkage on the unweighted Unifrac distance.

# ___________________________________________________________________________________________________________
# FIGURE 1A

# Get unweighted unifrac
set.seed(123)
unw.uniF <- UniFrac(physeq.rel, weighted=FALSE)

# Hierarchical clustering on Bray-Curtis distance
hc.average <- hclust(unw.uniF, method = "ward.D")

# Make into dendogram
dend.average <- as.dendrogram(hc.average)

# Function that will allow to color the leaves of the tree according to disease type (healthy vs IBS)
color_leaf<<-function(n){
    if(is.leaf(n)){
        # take the current attributes
        a=attributes(n)
        
        # deduce the line in the original data, to get the corresponding disease status
        line=match(attributes(n)$label, sample_names(physeq.fukui))
        disease=sample_data(physeq.fukui)[line, "host_disease"];
            if(disease=="Healthy"){col_disease="blue"};if(disease=="IBS"){col_disease="red"}
        
        #Modification of leaf attribute
        attr(n,"nodePar") <- c(a$nodePar, list(cex=0.8, pch=20, col=col_disease, # size & shape leaf
                                               lab.cex=0.4)) # size leaf label
        }
    return(n)
}

# Apply the color attributes to the dendograms
dend.average <- dendrapply(dend.average, color_leaf)
dend.average <- color_branches(dend.average, k=3)

# Plot
par(mar=c(3,2,3,7), xpd=TRUE)
plot(dend.average , main="Hierarchical clustering: Ward", horiz = T)
legend(x = 50,
     legend = c("Healthy" , "IBS"), 
     col = c("blue", "red"), 
     pch = c(20,20), bty = "n",  pt.cex = 1.5, cex = 1, 
     text.col = "black", horiz = FALSE, inset = c(0, 0.1))